Redefine Success
For three years, the standard complaint about Performance Max was that it was a black box you fed money into. That was a fair description until about twelve months ago. It isn't anymore.
Between the 2026 controls rollout and the broader shift toward AI-assisted account management, the question for serious advertisers has flipped. It's no longer "should I trust pMax?" — it's "am I actually using all the steering I now have, or am I still managing it like it's 2023?"
This blueprint is what I've learned running Performance Max across five accounts in 2025 and 2026 — three e-commerce, two lead-gen, ranging from $8K/month to $90K/month in spend. It's structured around three questions that determine whether your pMax campaign works:
Are you feeding the algorithm the right inputs? (Signals, assets, conversion data)
Are you using the new control surfaces Google shipped? (Negative keywords, placement exclusions, URL contains, audience exclusions)
Are you using AI in the parts of the workflow where it actually saves time? (Hint: it's not where most agencies are using it)
Skip the parts you already know. The Table of Contents is below.
Table of Contents
What Actually Changed in Performance Max (2025–2026)
The Input Layer: Signals, Assets, and Conversion Data
The Control Layer: The New Steering Wheel Google Finally Gave Us
The AI Layer: Where AI Genuinely Earns Its Keep in pMax
The Diagnostic Loop: How to Tell If Your pMax Is Actually Working
The September 2026 AI Max Migration: What's Coming
Putting It All Together: A 90-Day pMax Optimization Playbook
1. What Actually Changed in Performance Max (2025–2026)
If you stopped paying close attention to pMax updates in 2024, here's what you missed. These aren't incremental changes — they fundamentally alter the optimal strategy.
Campaign-level negative keywords expanded to 10,000 per campaign. This single change rewrites the playbook. Previously, advertisers had to rely on account-level negatives (capped at 1,000) or play games with brand exclusion lists. You can now build genuinely robust negative keyword strategies inside individual pMax campaigns. Note the limitation: negative keywords only affect Search and Shopping inventory inside pMax, not YouTube, Display, Gmail, or Discover. For those you still need placement and topic exclusions.
Placement exclusions at the account level. You can now block specific display placements and search partner sites from pMax across the entire account. This was previously the single biggest reason agencies refused to run pMax for clients with brand safety requirements. That objection is now mostly dead.
URL Contains rules for product feeds. Campaigns running with a Merchant Center feed can now target by URL pattern. Pages with "/new-release" in the URL can be one campaign, "/clearance" another, "/haircare" another. This is the closest pMax has come to Shopping campaign-level structural control.
First-party audience exclusions. You can now exclude your existing customer list from a pMax campaign. The use case is obvious in retrospect — separate acquisition from retention. Run one pMax for new customers (existing customers excluded) and a different campaign or budget for retention. The acquisition campaign no longer wastes spend re-converting people who would have come back anyway.
Budget projections inside pMax. You can now see end-of-month spend projections at the campaign level, and model how budget shifts affect performance. This is small but useful — particularly for accounts where pMax is one of several campaign types competing for total account budget.
Asset group A/B testing. You can now split traffic between two creative sets and measure performance. This is genuinely new — pMax previously offered no way to isolate creative impact.
15 videos per asset group (up from 5). YouTube and Demand Gen inventory expanded substantially. If you're not creating short-form video at scale, you're leaving inventory on the table.
Full audience reporting with demographics and segments. You can now see age, gender, and audience segment performance breakdowns. Previously this data was directional at best.
The aggregate effect of these changes: pMax in 2026 is a campaign type you can actually steer, not just a budget you assign to Google's machine and pray over. The strategies in the rest of this post assume you're operating with these controls available.
2. The Input Layer: Signals, Assets, and Conversion Data
Performance Max is a machine learning system, and like any ML system, the output quality is determined by input quality. The advertisers I've seen succeed with pMax all spend a disproportionate amount of time on inputs — and the ones who fail are almost always trying to fix output problems (bad placements, weak conversions) without fixing the inputs that caused them.
Conversion Data: The First Thing to Fix
Before anything else, audit your conversion setup. I have done this on every pMax account I've taken over in the last 18 months, and roughly 60% of them had a conversion tracking problem that was actively misleading the algorithm. Common ones.
Conversion actions tracking thank-you page views that fire on the customer dashboard (so every account login counts as a "conversion")
Multiple conversion actions all set to "primary," creating contradictory training signals
Lead-gen accounts optimizing for "form starts" instead of "form submits"
Values assigned wrong — every conversion worth $1, or every conversion worth the cart subtotal but not the order total
The pMax algorithm will faithfully optimize toward whatever you tell it to optimize for. If the signal is junk, the output is junk. There's no clever bidding strategy that fixes a bad conversion setup.
The 2026 update worth knowing about: Enhanced Conversions for Leads has matured significantly. If you're running lead-gen and not using EC for Leads with hashed email/phone pass-back, you're under-utilizing the platform by 20–30% in my experience. The algorithm gets significantly better data about which clicks actually became customers.
Audience Signals: The Most Misunderstood Input
Audience signals in pMax are not targeting. They are hints to the algorithm about what good prospects look like, used during the learning phase to find similar audiences. The algorithm will routinely serve ads outside the signaled audiences — and that's working as intended.
The mistake I see most often is treating signals like Display campaign audiences and getting upset when impressions show up outside them. The right mental model: audience signals are a starting point that the algorithm uses to bootstrap learning, then expands based on real conversion data.
What I've found works best in 2026:
One asset group, multiple signal layers. Stack 1st-party data (customer match), in-market segments, and custom segments based on competitor search behavior. The algorithm picks up the strongest signals.
Customer match lists with at least 1,000 hashed emails. Smaller lists get suppressed and provide weak training data.
Custom segments using URLs from your top-performing organic content. If your blog is ranking for "best [product category]", build a custom segment of people who visit those URLs. This is one of the highest-quality signals available.
Avoid demographic signals as starting inputs unless your offer is genuinely demographic. They tend to over-constrain learning during the early phase.
Assets: The Underrated Lever
Asset quality is the single biggest determinant of pMax performance in 2026, and most advertisers treat it as an afterthought. With 15 videos per asset group now allowed, and AI tools dramatically reducing video production costs, there is no longer an excuse for a thin creative library.
Minimum viable asset group in 2026:
15 headlines (mix of feature-led, benefit-led, and brand)
5 long headlines
5 descriptions
20 images at multiple aspect ratios (square, landscape, portrait)
5 logos
At least 5 short videos (6–15 seconds) — even AI-generated ones outperform no video at all
Final URL paths configured properly
If you're running pMax with the minimum (3 headlines, 1 video Google auto-generated for you), the campaign is operating on starvation rations. The algorithm has nothing to test.
Side note on a tool we built: This is part of why I built and now sell a Creative Asset Performance Analyzer script — it pulls asset-level performance from pMax (using the new 2026 asset reporting endpoints), identifies which combinations are actually serving impressions and converting, and flags assets the algorithm has stopped serving. Most pMax campaigns have 20–40% of their assets sitting dormant. If that sounds useful, it's in the Adstronaut script library. Back to the post.
3. The Control Layer: The New Steering Wheel Google Finally Gave Us
This is the section that's most different from any pMax post written before mid-2025. The controls listed here are all new in the last 12 months, and most of them are still under-utilized by the advertisers I audit.
Campaign-Level Negative Keywords (the 10,000 cap)
The strategy I use, in order of priority.
Tier 1 — Universal exclusions (apply to every pMax campaign):
Job-seeker terms: "jobs", "careers", "hiring", "salary", "employment"
Free-seeker terms: "free", "freeware", "no cost", "diy"
Educational intent: "how to", "tutorial", "guide" (unless your offer is educational)
Competitor brand names you specifically don't want association with
Tier 2 — Account-specific exclusions:
Past conversion search terms that converted poorly (high refund, low LTV)
Terms that consistently produced bot traffic in past Search campaigns
Geographic terms outside your service area for local businesses
Tier 3 — Pulled from Search Terms Report (review weekly):
The Search Terms Report inside pMax is now reasonably useful. Review it weekly. Add any term with 50+ impressions and 0 conversions to negatives. Add any term that's clearly irrelevant regardless of conversion count.
The 10,000 cap means you don't need to be precious. Be aggressive. The cost of an unnecessary negative is small; the cost of bleeding budget on irrelevant terms is large.
Placement Exclusions
Account-level placement exclusions are now your brand safety layer. The exclusions I apply by default on every account:
App categories irrelevant to the business (games for B2B, etc.)
Known low-quality placement domains (the agency community maintains shared lists; ask in your peer groups)
YouTube channels you've identified as low-quality from past Display campaigns
Search partner sites if your past Search campaign data shows partner sites converting at <25% of Google.com performance
This is also where the placement report inside pMax becomes essential. Pull it monthly. The signal is noisy on small accounts (you'll see weird placements that got 2 impressions and didn't convert), but on accounts spending $20K+/month you'll find patterns worth excluding.
URL Contains (for E-commerce)
If you're running an e-commerce account with a Merchant Center feed, URL Contains rules are the closest thing to old-school campaign structure that pMax allows. The strategy:
Separate asset groups (or separate campaigns) by URL path patterns that signal different intent and margin
New arrivals vs. evergreen products often have different conversion patterns; split them
Sale/clearance pages typically have higher CVR but lower margin — isolate them so you can target ROAS differently
High-margin product categories should get their own structure with more aggressive bid targets
This brings back something resembling the segmentation discipline you had with Standard Shopping campaigns, before pMax effectively forced everyone into one bucket.
First-Party Audience Exclusions: The Acquisition/Retention Split
This is the highest-leverage 2026 update for accounts with meaningful customer lists. The setup:
Campaign A: New Customer Acquisition. Existing customer match list excluded. Target CPA is calculated based on first-purchase margin only.
Campaign B: Retention/Repeat. Limited to the existing customer match list (use it as a positive signal, not exclusion). Target CPA can be higher because LTV-based math applies.
For accounts I've migrated to this structure, the result is usually 15–30% more new customers at the same blended CAC, because the acquisition campaign is no longer subsidized by easy repeat conversions that would have happened anyway.
4. The AI Layer: Where AI Genuinely Earns Its Keep in pMax
Most "AI for Google Ads" content focuses on the obvious wins — AI-generated headlines, AI-generated images. Those are fine, but they're the lowest-leverage application of AI in a pMax workflow. The higher-leverage uses are in analysis, not generation.
Here's where I've found AI tools (Claude and ChatGPT primarily) save real time in the pMax workflow:
Search Terms Analysis at Scale
The pMax Search Terms Report can return thousands of terms on a large account. Manually categorizing them — relevant, irrelevant, exclude as negative, possibly relevant — is hours of work per account per week.
What works: export the Search Terms Report as CSV, paste it into an AI tool with a prompt like "Categorize each of these search terms into one of: relevant_high_intent, relevant_low_intent, irrelevant_should_exclude, ambiguous_needs_review. Return as a CSV with the categorization column added."
A 30-minute task becomes a 3-minute task. The AI gets it right on roughly 85% of terms; you review the ambiguous bucket and the ones where AI categorization seems off.
Asset Group Naming and Theming
When you have 6 asset groups across 3 pMax campaigns, naming discipline matters or you lose track of what's testing what. AI is useful for generating consistent naming conventions and for translating fuzzy thematic ideas ("the customers who care about durability") into asset group names that make sense six months from now.
Reporting and Analysis
This is where AI saves the most time in my workflow. Pulling pMax performance data into a sheet, then asking an AI to summarize what happened week-over-week, identify anomalies, and suggest hypotheses for the changes — this is genuinely faster than doing it manually. Not because the AI is smarter than I am, but because writing the summary is the slow part, and AI is fast at writing.
I use this weekly. Roughly 90 minutes of analysis work becomes 25 minutes. The AI doesn't replace the judgment about what to do next; it replaces the typing.
Asset Generation (Use Sparingly)
AI-generated images and copy for pMax assets work, but with caveats. Image generation is now good enough for product photography backgrounds, lifestyle scenes, and abstract brand imagery. It's still not good enough for anything where brand consistency matters or where text appears in the image (text generation in AI images remains unreliable as of mid-2026).
AI-generated headlines work, but the algorithm-generated variations Google produces from your existing headlines are often as good or better. Where AI headline generation earns its keep is in expanding into adjacent angles — taking your existing 8 headlines and producing 7 more that approach the same offer from different psychological angles.
What AI Is Not Good For in pMax
Worth flagging the limitations honestly:
AI cannot make strategic decisions about your account. It can pattern-match to common cases, but it doesn't understand your margins, your competitive position, or your specific business context. Treat its strategic suggestions as starting points to evaluate, not recommendations to execute.
AI cannot reliably predict which assets will perform. The space of "what works in ads" is too narrow and brand-dependent for current models to generalize well.
AI cannot replace the judgment about when to give a campaign more time vs. when to restructure. This requires experience reading the specific signs of an account that's learning vs. an account that's stuck.
5. The Diagnostic Loop: How to Tell If Your pMax Is Actually Working
The hardest skill in pMax management is reading the signals correctly. Here's what I look at, in priority order, for any pMax account.
Weekly cadence:
Conversion volume trend. Smooth growth or stability is good. Sudden drops or spikes warrant investigation.
Search Terms Report. Look for irrelevant terms, add to negatives.
Asset performance ratings. Anything marked "Low" should be replaced within 30 days. The "Best" rated assets should be expanded into variations.
Channel breakdown report. Watch for shifts. If Search inventory share is dropping and YouTube/Display share is rising, the algorithm is finding cheaper conversions in lower-intent inventory — sometimes good, sometimes a sign of declining intent quality.
Monthly cadence:
Placement report. Identify and exclude poor placements.
Audience breakdown. Look for demographics or segments significantly underperforming or overperforming. The new 2026 reporting makes this far more useful than it used to be.
New customer acquisition rate. Critical if you've split acquisition and retention campaigns.
Asset group performance. Are your asset groups differentiating? If not, the theming isn't doing its job.
Quarterly cadence:
Bidding strategy review. Is your target CPA or ROAS still calibrated to current margins? Most accounts let this drift.
Conversion model review. Are you optimizing for the right action? Customer feedback or business priorities may have shifted.
Structural review. Should asset groups be split or merged? Should you add or close pMax campaigns?
The single most important indicator: conversion volume holding or growing at a stable CPA/ROAS over rolling 28-day periods. Day-to-day noise is normal. Week-to-week noise is normal. Month-to-month trend is what matters. If the rolling 28-day numbers are stable and you have growth ambition, increase budgets gradually (15–25% increments) and watch the next 28 days. If they're degrading, work the diagnostic checklist above.
6. The September 2026 AI Max Migration: What's Coming
Worth knowing about: Google announced in April 2026 that Dynamic Search Ads will begin auto-upgrading to AI Max starting in September 2026. AI Max is being positioned as the new defaults for Search campaigns broadly, and pMax sits inside this trajectory.
What this likely means for pMax operators:
More aggressive automation of asset generation, with less explicit control over what the algorithm produces
Tighter integration between Search and pMax, blurring the lines between campaign types
Continued movement toward "input quality" as the primary advertiser lever, with less emphasis on structural control
The strategic implication: invest in the inputs (conversion data, customer match lists, custom segments, asset libraries) and in the analysis layer. Structural campaign management as a skill is gradually depreciating; input quality and diagnostic skill are appreciating.
This is exactly why I lean on scripts and automation for the diagnostic and optimization workflows — the bottleneck is moving away from "do you know how to structure campaigns" toward "can you analyze account performance at scale and act on it quickly." Scripts and AI tooling both compound here.
7. Putting It All Together: A 90-Day pMax Optimization Playbook
For an account you've just taken over, or a pMax campaign you suspect is underperforming, here's the order I work through:
Days 1–14: Audit and Foundation
Audit conversion tracking. Fix anything broken. Don't touch the campaign until tracking is right.
Verify Enhanced Conversions are set up. For lead-gen, set up EC for Leads with hashed email/phone.
Build the audience signal stack: customer match list (1K+ contacts), top organic URL custom segments, in-market segments aligned to offer.
Build the asset library: minimum 15 headlines, 20 images, 5 videos per asset group.
Set up the acquisition/retention split if you have a customer list big enough to support it.
Days 15–30: Apply Controls
Build the negative keyword list (start with universal exclusions, add account-specific).
Apply account-level placement exclusions for known low-quality inventory.
If e-commerce, set up URL Contains rules to segment by product category or margin.
Configure brand exclusions if you're also running branded search separately.
Days 31–60: Test and Iterate
Set up asset group A/B test with at least one creative variation.
Weekly: review Search Terms Report, add negatives.
Bi-weekly: review asset performance, replace "Low" rated assets.
End of week 8: full month-over-month performance review.
Days 61–90: Optimize and Scale
Adjust target CPA/ROAS based on observed performance.
If performance is stable, begin budget scaling in 20% increments every 2 weeks.
Begin building the second asset group or campaign if expansion makes sense.
Document what worked and what didn't for the next account.
Final Thoughts
Performance Max in 2026 is a different campaign type than the one most playbooks were written for. The black-box reputation no longer fits. The new controls are real. The reporting is meaningfully better. The AI tools available to advertisers are dramatically more capable than they were even 18 months ago.
What hasn't changed is the underlying discipline: high-quality conversion data, deep audience signal libraries, deep creative libraries, ruthless analysis of where money is going and what's coming back. Advertisers who do these things win. Advertisers who don't, lose money to Google's defaults regardless of how good the platform's automation gets.
The advertisers who win biggest in 2026 will be the ones who pair sound fundamentals with the new control surfaces and with AI tooling that lets them analyze and act at scale. None of it is glamorous. All of it compounds.
